基于t混合模型的脑部核磁共振图像分割方法研究

发布时间：2018-04-20 15:49

本文选题：脑MR图像分割 + SMM　；参考：《华侨大学》2017年硕士论文

【摘要】：近几十年,医学影像技术飞速发展,特别是核磁共振成像技术(Magnetic Resonance Imaging,MRI)。该技术可以检测人体各个部位的疾病,应用极其广泛。医学图像分析为医生提供有效的参考信息,而医学图像分割是医学图像分析的基础工作。大脑对于人体至关重要,它直接控制着人体的运动、感觉以及实现人体的很多高级功能。因此,脑MR图像分割在医学领域尤为重要。而脑MR图像的高噪声、灰度不均匀、边缘模糊等特点为其分割带来了挑战。另外,专家参与图像分割的方式消耗时间长,而图像数量巨大,因此急需实现脑MR图像分割的自动化。基于混合模型的图像分割方法借助图像的灰度直方图,将图像中像素的灰度值拟合成某些特定的分布。其中,高斯混合模型(Gaussian Mixture Model,GMM)具有直观、实现简便、符合很多数据的分布等特点,因而得到了广泛应用。而尾部较高的学生t分布代替高斯分布后,混合模型在图像分割过程中对噪声更鲁棒,因此t混合模型(Student’s t Mixture Model,SMM)又引起了许多学者的关注。为了应对以上脑MR图像分割中的面对的挑战,本文在SMM的基础上做出了两点改进,并将改进后的模型用于脑MR图像分割:1)基于一种新的马尔可夫随机场(Markov Random Field,MRF)的表示方法,设计并实现了一种结合MRF和SMM的脑MR图像分割方法。该方法利用MRF将图像中像素的空间关系信息引入混合模型,因此得到的分割结果对噪声更鲁棒。而且本文推导出该模型求解的期望最大化算法(Expectation Maximization Algorithm,EM算法),该算法能直接得到模型中参数的封闭解,避免了复杂的计算操作。实验表明该方法的表现优于其它流行的脑MR图像分割方法及其它同样基于MRF的混合模型方法。2)针对传统的SMM用于图像分割时需要人为设定分割数目的缺点,本文提出一种基于截断狄利克雷过程的无限t混合模型(Truncated Dirichlet Process Student’s t Mixture Model,tDPSMM)的脑MR图像分割方法。该方法用截断狄利克雷过程作为模型的先验信息,达到了使模型自动确定分割数目的目标,而且保证了该无限混合模型推理、计算的简便性。本文也推导出该模型下的EM算法,最终利用贝叶斯最大后验概率准则将模型应用于脑MR图像分割。与其它无限混合模型方法及经典的脑MR图像分割方法进行对比实验后,我们发现该方法能更快速准确地实现脑MR图像分割。
[Abstract]:In recent decades, medical imaging technology has developed rapidly, especially magnetic Resonance imaging. This technique can detect diseases in various parts of human body and is widely used. Medical image analysis provides effective reference information for doctors, and medical image segmentation is the basic work of medical image analysis. The brain is very important to the human body. It directly controls the movement, sensation and many advanced functions of the human body. Therefore, brain Mr image segmentation is particularly important in the field of medicine. However, the high noise, uneven grayscale and edge blur of Mr images bring challenges to the segmentation of Mr images. In addition, the way that experts participate in image segmentation takes a long time, and the number of images is huge, so it is urgent to realize the automation of brain Mr image segmentation. The method of image segmentation based on mixed model is based on the gray histogram of the image, and the gray value of the pixel in the image is combined to some special distribution. Among them, Gao Si mixed model Gaussian Mixture Model-GMMM has been widely used because of its intuitionistic, simple implementation and consistent with the distribution of many data. When the student t distribution with higher tail is replaced by Gao Si distribution, the mixed model is more robust to noise in the image segmentation process, so the t mixed model (Students t Mixture Model SMMM) has attracted many scholars' attention. In order to meet the challenges in brain Mr image segmentation, two improvements are made on the basis of SMM, and the improved model is applied to segment brain Mr image at 1: 1) based on a new representation method of Markov random field Markov Random field (MRF). A brain Mr image segmentation method combining MRF and SMM is designed and implemented. In this method, the spatial relation information of pixels in the image is introduced into the mixed model by MRF, so the segmentation result is more robust to noise. Moreover, this paper deduces the expectation maximization Maximization algorithm EM algorithm, which can directly obtain the closed solution of the parameters in the model, and avoid the complicated calculation operation. Experiments show that the performance of this method is better than that of other popular brain Mr image segmentation methods and other mixed model methods based on MRF. In this paper, a method of brain Mr image segmentation based on truncated Dirichlet Process Student's t Mixture Model DPS MMM is proposed. The method uses truncated Delikley process as the prior information of the model, and achieves the goal of automatically determining the number of segments of the model, and ensures the simplicity of reasoning and calculation of the infinite mixed model. The EM algorithm based on this model is also derived in this paper. Finally, the model is applied to the segmentation of brain Mr image by using Bayesian maximum a posteriori probability criterion. Compared with other infinite mixed model methods and classical brain Mr image segmentation methods, we find that this method can achieve the segmentation of brain Mr images more quickly and accurately.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R445.2;TP391.41

【相似文献】